minisql

A tiny SQL engine in C — one file, no dependencies. It loads CSV files as tables and runs SELECT queries over them, using the same parse → plan → execute pipeline a real database uses, small enough to read start to finish.

Build & run

make
./minisql          # interactive prompt; type .help for commands

Example

./minisql <<'SQL'
.load data/albums.csv albums
SELECT genre, COUNT(*) FROM albums GROUP BY genre ORDER BY 2 DESC;
SQL

What it supports

• SELECT with WHERE, INNER/LEFT JOIN, GROUP BY/HAVING, ORDER BY, LIMIT, DISTINCT
• Expressions (+ - * /, AND/OR/NOT, LIKE, IN, BETWEEN, IS NULL), functions (UPPER, LENGTH, ROUND, COALESCE, …), and aggregates (COUNT/SUM/AVG/MIN/MAX, COUNT(DISTINCT))
• A small query planner: hash joins, join reordering, predicate pushdown, and optional hash/sorted indexes (.index)
• Errors on unknown columns instead of quietly returning nothing

Semantics worth knowing

Where dialects disagree, minisql errs on the side of failing loudly:

• SELECT aliases are visible only in ORDER BY (like PostgreSQL / standard SQL; sqlite also allows them in WHERE/HAVING). Using one elsewhere is a bind error, never a silent NULL.
• ORDER BY must name a selected column or position — sorting happens after projection, so anything else errors instead of being ignored.
• Integer arithmetic (+ - *, SUM) that overflows 64 bits promotes to double instead of wrapping (sqlite raises an error here).
• Ragged CSV rows: missing fields become NULL; extra fields are dropped with a warning.
• The REPL is line-based: each statement must fit on one line (no ; continuation across lines).
• Types are strict: comparing a string column to a number matches nothing (sqlite would coerce via column affinity — minisql never converts silently).
• Strings are bytes: LENGTH and LIKE's _ count bytes, and UPPER/LOWER fold ASCII only, so multi-byte UTF-8 characters count per byte.

How it's built

CSV → tokenizer → parser (AST) → binder → planner → executor

Everything lives in main.c (~2,700 lines), in top-to-bottom order.

Tests

make test      # built-in checks + regression suite (every past bug stays fixed)
make stress    # real queries, cross-checked against sqlite3 (needs sqlite3)